Cursor · Phone screen
In process2026 Cursor Technical Phone Screen, Role and Hiring Level Not Reported: 1 Experience
Cursor interview reports
- New Grad: one full loop from November 2025: a chat with the hiring manager, a 60-minute hash tree round inside a repo, and a day of building a feature in Cursor's own code.
- Posted after the SpaceX acquisition announcement: seven reports from July to October 2026: six experienced candidates (five software engineers and one machine learning engineer) and one software engineer whose hiring level wasn't given.
- Role and hiring level not reported (current): one technical phone screen from January 2026, a two-part per-user rate limiter.
Coverage: One candidate in the United States. The post didn't name the role or say whether it was new grad or experienced. One technical phone screen, posted January 20, 2026. The candidate passed and moved to the next round.
What came up
The whole screen went to one task, a rate limiter for Cursor's API endpoints. Part 1 was a single per-user limit; Part 2 turned it into an interface that can stack several kinds of limits.
Questions we found
Algorithms / Data Structures
| Question | Frequency | All last seen dates |
|---|---|---|
| Completions API Rate Limiter (count limits plus a byte-size limit, all checked on each request) | 1 report | Report date: 2026-01-20 |
1 detailed candidate timeline
01: From one endpoint limit to five rules checked together
Report date: January 20, 2026 · Role: Not named (hiring level not reported) · Location: United States · Result: Passed, moved to the next round
Technical phone screen
Cursor's completions API and other endpoints are getting heavy use, so the team needs a rate limiter. Requests sit in memory as records of [user_id, request_url, body_size (bytes), timestamp (seconds)]. Every limit is tracked per user.
-
Part 1, one endpoint. Allow at most 3 requests per second to
v1/completions/standard. Denied requests don't count toward the limit. Other endpoints have no limit yet. -
Part 2, a rate limiter interface. Ten other developers now want limits on their endpoints, so the design has to support all of these at once:
v1/completions/chunked: at most 7 requests per minutev2/completions/chunked: at most 5 requests per secondv1/completions/chunked: allowed only if the total request size in the past minute stays within 100 bytes- at most 7 requests per hour across all endpoints
- the Part 1 rule
A request goes through only if every check passes. If any check fails, the request is dropped and counts toward nothing.
The candidate found Part 1 straightforward: fill in the given stub the simple way. Part 2 took more thought. They built an interface with two concrete classes to cover the different kinds of limits and got about 95% of the way there. The building blocks were done, but they ran out of time before setting up every limit in the main function. The interviewer wasn't too worried about that. The interviewer had said the goal was working code, not tuned time and space complexity, and cared most about how the main function would run every limiter on a request and decide to allow or deny it.
The candidate heard back about an hour after the call. Their main takeaway: keep thinking out loud, and bring the interviewer in as a stakeholder.
More Cursor interview reports
- New Grad
- Posted after the SpaceX acquisition announcement
- Role and hiring level not reported (current)
Comments
0No comments yet — be the first to share what you know.